The utilization of wind Lidar is beneficial for predicting rotor equivalent wind speed (REWS), which is indispensable for planning and optimizing wind power generation systems. Nevertheless, simulating the Lidar measurements and unraveling the complex mapping between REWS and Lidar data continues to pose a significant challenge. Starting on Lidar measurement modeling, a novel prediction framework is proposed for REWS robust prediction. Specifically, the Lidar measurement data is decomposed by the variational mode decomposition (VMD), and the resulted components are reconstructed through weight recognition using correlation Euclidean distance (CED) method. On this basis, a novel splGRU prediction model is constructed, in which the self-paced learning (SPL) strategy is introduced to combine the gate recurrent unit (GRU) network. Therein, the weights of training samples are dynamically adjusted during training, enhancing the GRU model adaptation and learning capabilities for hard samples. To verify the effectiveness of the proposed method, TurbSim and Matlab software are utilized to simulate and generate two sets of Lidar measurements with different mean wind speeds and turbulence intensities. The experimental results indicate that, compared to the traditional GRU model, the splGRU model improves normalized root mean square error by 5.8 % and 7.6 % in datasets 1 and 2, respectively; the robustness metric, 90th percentile minimax regret (regret90th), is improved by 19.5 % and 9.8 %, respectively. Meanwhile, the splGRU also shows better prediction accuracy and robustness when comparing the advanced splLSTM model. The proposed VMD-CED-splGRU method holds promise for future integration into model predictive control systems, which could achieve super control performance of large-scale wind turbines